A developer known as stmonty has trained a small AI model on a single RTX 3080 Ti to play Pokémon Red, according to a blog post covered by Tom's Hardware. The model, built on the LeWorldModel research co-authored by Yann LeCun, uses only 12.5 million parameters and was trained locally on consumer hardware. Rather than predicting entire screens, it works from compressed summaries of just 192 numbers, learning what each button press does by observing the resulting display.
The training data consisted of 42,382 grayscale frames from over 1,000 short runs in Professor Oak's lab, including scripted routes, those routes with random button presses, and random wandering. The developer noted that mixing in random presses was necessary because a model trained only on clean runs might never learn what the B button does. The initial training is reward-free, with goals introduced later during planning, when the model searches through 512 plan samples per round and keeps one in eight.
After fine-tuning, the model successfully selected Squirtle, and across 100 runs from the same save it obtained a starter 52 times, compared with zero for random button presses and one for an untrained model. The developer described the project as a way to learn and have fun, and acknowledged that simply scaling up the model likely would not lead to beating the game. The code is open source on GitHub as lePokeRed, with a CUDA GPU recommended.